Capstone: Adapt

The capstone runs the complete methodology on a task you choose — ideally from your real work (sanitized per your data policy), alternatively a course-provided task menu (brand-voice product descriptions, structured extraction from a gnarly document type, a routing classifier with asymmetric stakes). The certification attests the method: gate, data, runs, eval, economics, ship judgment.

1. 1Pass the gate for real: eval set built (100+ cases, held-out slice sealed), best-prompt ceiling established on your chosen student with genuine effort (reviewers check for strawman prompts — three logged iteration rounds minimum), the measured gap and the adaptation brief with exit criteria. If your gap evaporates under honest prompting, switching projects to that finding is a permitted and respected capstone: document the ceiling work and the decision memo; you've saved a fictional team a real quarter.
2. 2Build the dataset with the full Module 2 discipline: labeling guide, curriculum-proportioned coverage, cleaning pipeline as code, contamination check against your eval, PII/rights hygiene appropriate to your source, the stratified audit with agreement score, and the versioned manifest. Reviewers weight this phase heaviest — because it predicts everything else.
3. 3Run adaptation with the protocol: wire-test on the record, one clean SFT run with frozen inputs, checkpoint selection by validation, full eval with the three-column comparison and by-class error analysis, then at least one data-driven improvement loop (the v1.1 retrain) — and a preference pass only if your error analysis shows a preference-shaped gap (justify it either way in one paragraph; 'no DPO needed, here's why' is a fine answer and a common correct one).
4. 4Keep the experiment log as a first-class artifact: every run — config, dataset version, scores, cost, decision taken. The log tells the project's story better than the report will; submissions with reconstructed-after-the-fact logs read exactly like what they are.